def glm_doit(self, csvFilename, csvPathname, timeoutSecs=30):
print "\nStarting parse of", csvFilename
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname, key2=csvFilename + ".hex", timeoutSecs=10)
y = "10"
x = ""
# NOTE: hastie has two values, -1 and 1. To make H2O work if two valued and not 0,1 have
kwargs = {
'x': x, 'y': y, 'case': '1', 'destination_key': 'gg',
# better classifier it flipped? (better AUC?)
'max_iter': 10,
'case': -1, 'case_mode': '=',
'num_cross_validation_folds': 0,
'lambda': '1e-8,1e-4,1e-3',
'alpha': '0,0.25,0.8',
# hardwire threshold to 0.5 because the dataset is so senstive right around threshold
# otherwise, GLMGrid will pick a model with zero coefficients, if it has the best AUC
# to avoid my checker complaining about all zero coefficients, force the threshold to 0.5
'thresholds': '0.5',
# 'thresholds': '0.2:0.8:0.1'
}
start = time.time()
print "\nStarting GLMGrid of", csvFilename
glmGridResult = h2o_cmd.runGLMGridOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
print "GLMGrid in", (time.time() - start), "secs (python)"
# still get zero coeffs..best model is AUC = 0.5 with intercept only.
h2o_glm.simpleCheckGLMGrid(self,glmGridResult, allowZeroCoeff=True,**kwargs)
def test_GLMGrid_basic_benign(self):
csvFilename = "benign.csv"
print "\nStarting", csvFilename
csvPathname = h2o.find_file('smalldata/logreg' + '/' + csvFilename)
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname, key2=csvFilename + ".hex")
# columns start at 0
# cols 0-13. 3 is output
# no member id in this one
y = "3"
x = range(14)
x.remove(0) # 0. skipping causes coefficient of 0 when used alone
x.remove(3) # 3 is output
x = ','.join(map(str, x))
# just run the test with all x, not the intermediate results
print "\nx:", x
print "y:", y
kwargs = {
'x': x, 'y': y, 'n_folds': 0,
'lambda': '1e-8:1e-2:100',
'alpha': '0,0.5,1',
'thresholds': '0:1:0.01'
}
# fails with n_folds
print "Not doing n_folds with benign. Fails with 'unable to solve?'"
gg = h2o_cmd.runGLMGridOnly(parseKey=parseKey, timeoutSecs=120, **kwargs)
# check the first in the models list. It should be the best
colNames = [ 'STR','OBS','AGMT','FNDX','HIGD','DEG','CHK',
'AGP1','AGMN','NLV','LIV','WT','AGLP','MST' ]
h2o_glm.simpleCheckGLMGrid(self, gg, None, **kwargs)
def test_GLMGrid_basic_prostate(self):
csvFilename = "prostate.csv"
print "\nStarting", csvFilename
# columns start at 0
csvPathname = h2o.find_file('smalldata/logreg' + '/' + csvFilename)
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname, key2=csvFilename + ".hex")
y = "1"
x = range(9)
x.remove(0) # 0. member ID. not used.
x.remove(1) # 1 is output
x = ','.join(map(str, x))
# just run the test with all x, not the intermediate results
print "\nx:", x
print "y:", y
# FIX! thresholds is used in GLMGrid. threshold is used in GLM
# comma separated means use discrete values
# colon separated is min/max/step
# FIX! have to update other GLMGrid tests
kwargs = {
'x': x, 'y': y, 'n_folds': 2,
'beta_epsilon': 1e-4,
'lambda': '1e-8:1e3:100',
'alpha': '0,0.5,1',
'thresholds': '0:1:0.01'
}
gg = h2o_cmd.runGLMGridOnly(parseKey=parseKey, timeoutSecs=120, **kwargs)
colNames = ['D','CAPSULE','AGE','RACE','DPROS','DCAPS','PSA','VOL','GLEASON']
# h2o_glm.simpleCheckGLMGrid(self, gg, colNames[xList[0]], **kwargs)
h2o_glm.simpleCheckGLMGrid(self, gg, None, **kwargs)
def glm_doit(self, csvFilename, csvPathname, timeoutSecs=30):
print "\nStarting parse of", csvFilename
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname, key2=csvFilename, timeoutSecs=10)
y = "10"
x = ""
# NOTE: hastie has two values, -1 and 1. To make H2O work if two valued and not 0,1 have
kwargs = {
"x": x,
"y": y,
"case": "1",
"destination_key": "gg",
# better classifier it flipped? (better AUC?)
"max_iter": 10,
"case": -1,
"case_mode": "=",
"num_cross_validation_folds": 0,
"lambda": "1e-8,1e-4,1e-3",
"alpha": "0,0.25,0.8",
"thresholds": "0.2:0.8:0.1",
}
start = time.time()
print "\nStarting GLMGrid of", csvFilename
glmGridResult = h2o_cmd.runGLMGridOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
print "GLMGrid in", (time.time() - start), "secs (python)"
h2o_glm.simpleCheckGLMGrid(self, glmGridResult, **kwargs)
def test_GLMGrid_basic_prostate(self):
csvFilename = "prostate.csv"
print "\nStarting", csvFilename
# columns start at 0
csvPathname = h2o.find_file('smalldata/logreg' + '/' + csvFilename)
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname,
key2=csvFilename + ".hex")
y = "1"
x = range(9)
x.remove(0) # 0. member ID. not used.
x.remove(1) # 1 is output
x = ','.join(map(str, x))
# just run the test with all x, not the intermediate results
print "\nx:", x
print "y:", y
# FIX! thresholds is used in GLMGrid. threshold is used in GLM
# comma separated means use discrete values
# colon separated is min/max/step
# FIX! have to update other GLMGrid tests
kwargs = {
'x': x,
'y': y,
'n_folds': 2,
'beta_epsilon': 1e-4,
'lambda': '1e-8:1e3:100',
'alpha': '0,0.5,1',
'thresholds': '0:1:0.01'
}
gg = h2o_cmd.runGLMGridOnly(parseKey=parseKey,
timeoutSecs=120,
**kwargs)
colNames = [
'D', 'CAPSULE', 'AGE', 'RACE', 'DPROS', 'DCAPS', 'PSA', 'VOL',
'GLEASON'
]
# h2o_glm.simpleCheckGLMGrid(self, gg, colNames[xList[0]], **kwargs)
h2o_glm.simpleCheckGLMGrid(self, gg, None, **kwargs)
def test_GLMGrid_basic_benign(self):
csvFilename = "benign.csv"
print "\nStarting", csvFilename
csvPathname = h2o.find_file('smalldata/logreg' + '/' + csvFilename)
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname,
key2=csvFilename + ".hex")
# columns start at 0
# cols 0-13. 3 is output
# no member id in this one
y = "3"
x = range(14)
x.remove(0) # 0. skipping causes coefficient of 0 when used alone
x.remove(3) # 3 is output
x = ','.join(map(str, x))
# just run the test with all x, not the intermediate results
print "\nx:", x
print "y:", y
kwargs = {
'x': x,
'y': y,
'n_folds': 0,
'lambda': '1e-8:1e-2:100',
'alpha': '0,0.5,1',
'thresholds': '0:1:0.01'
}
# fails with n_folds
print "Not doing n_folds with benign. Fails with 'unable to solve?'"
gg = h2o_cmd.runGLMGridOnly(parseKey=parseKey,
timeoutSecs=120,
**kwargs)
# check the first in the models list. It should be the best
colNames = [
'STR', 'OBS', 'AGMT', 'FNDX', 'HIGD', 'DEG', 'CHK', 'AGP1', 'AGMN',
'NLV', 'LIV', 'WT', 'AGLP', 'MST'
]
h2o_glm.simpleCheckGLMGrid(self, gg, None, **kwargs)
def test_B_benign(self):
csvFilename = "benign.csv"
print "\nStarting", csvFilename
csvPathname = h2o.find_file('smalldata/logreg' + '/' + csvFilename)
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname, key2=csvFilename + ".hex")
# columns start at 0
# cols 0-13. 3 is output
# no member id in this one
y = "3"
xList = []
for appendx in xrange(14):
if (appendx == 0):
print "\nSkipping 0. Causes coefficient of 0 when used alone"
elif (appendx == 3):
print "\n3 is output."
else:
xList.append(appendx)
x = ','.join(map(str, xList))
# just run the test with all x, not the intermediate results
print "\nx:", x
print "y:", y
kwargs = {
'x': x, 'y': y, 'num_cross_validation_folds': 0,
'lambda': '1e-8:1e-2:100',
'alpha': '0,0.5,1',
'thresholds': '0:1:0.01'
}
# fails with num_cross_validation_folds
print "Not doing num_cross_validation_folds with benign. Fails with 'unable to solve?'"
gg = h2o_cmd.runGLMGridOnly(parseKey=parseKey, timeoutSecs=120, **kwargs)
# check the first in the models list. It should be the best
colNames = [ 'STR','OBS','AGMT','FNDX','HIGD','DEG','CHK',
'AGP1','AGMN','NLV','LIV','WT','AGLP','MST' ]
# h2o_glm.simpleCheckGLMGrid(self, gg, colNames[xList[-1]], **kwargs)
h2o_glm.simpleCheckGLMGrid(self, gg, None, **kwargs)
def glm_doit(self, csvFilename, csvPathname, timeoutSecs=30):
print "\nStarting parse of", csvFilename
parseKey = h2o_cmd.parseFile(csvPathname=csvPathname,
key2=csvFilename + ".hex",
timeoutSecs=10)
y = "10"
x = ""
# NOTE: hastie has two values, -1 and 1. To make H2O work if two valued and not 0,1 have
kwargs = {
'x': x,
'y': y,
'case': '1',
'destination_key': 'gg',
# better classifier it flipped? (better AUC?)
'max_iter': 10,
'case': -1,
'case_mode': '=',
'n_folds': 2,
'lambda': '1e-8,1e-4,1e-3',
'alpha': '0,0.25,0.8',
# hardwire threshold to 0.5 because the dataset is so senstive right around threshold
# otherwise, GLMGrid will pick a model with zero coefficients, if it has the best AUC
# to avoid my checker complaining about all zero coefficients, force the threshold to 0.5
'thresholds': '0.5',
# 'thresholds': '0.2:0.8:0.1'
}
start = time.time()
print "\nStarting GLMGrid of", csvFilename
glmGridResult = h2o_cmd.runGLMGridOnly(parseKey=parseKey,
timeoutSecs=timeoutSecs,
**kwargs)
print "GLMGrid in", (time.time() - start), "secs (python)"
# still get zero coeffs..best model is AUC = 0.5 with intercept only.
h2o_glm.simpleCheckGLMGrid(self,
glmGridResult,
allowZeroCoeff=True,
**kwargs)
def test_GLM_convergence_1(self):
SYNDATASETS_DIR = h2o.make_syn_dir()
tryList = [
(100, 50, 'cD', 300),
(100, 100, 'cE', 300),
(100, 200, 'cF', 300),
(100, 300, 'cG', 300),
(100, 400, 'cH', 300),
(100, 500, 'cI', 300),
]
### h2b.browseTheCloud()
lenNodes = len(h2o.nodes)
USEKNOWNFAILURE = True
for (rowCount, colCount, key2, timeoutSecs) in tryList:
SEEDPERFILE = random.randint(0, sys.maxint)
csvFilename = 'syn_%s_%sx%s.csv' % (SEEDPERFILE,rowCount,colCount)
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "\nCreating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
if USEKNOWNFAILURE:
csvFilename = 'failtoconverge_100x50.csv'
csvPathname = h2o.find_file('smalldata/logreg/' + csvFilename)
parseKey = h2o_cmd.parseFile(None, csvPathname, key2=key2, timeoutSecs=10)
print csvFilename, 'parse time:', parseKey['response']['time']
print "Parse result['destination_key']:", parseKey['destination_key']
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
print "\n" + csvFilename
y = colCount
kwargs = {
'max_iter': 10,
'weight': 1.0,
'link': 'familyDefault',
'n_folds': 2,
'beta_epsilon': 1e-4,
#***********
'lambda': '1e-8:1e-3:1e2',
'alpha': '0,0.5,.75',
'thresholds': '0,1,0.2'
}
if USEKNOWNFAILURE:
kwargs['y'] = 50
else:
kwargs['y'] = y
emsg = None
for i in range(25):
start = time.time()
glm = h2o_cmd.runGLMGridOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, noise=("Jstack", None), **kwargs)
print 'glm #', i, 'end on', csvPathname, 'took', time.time() - start, 'seconds'
# we can pass the warning, without stopping in the test, so we can
# redo it in the browser for comparison
warnings = h2o_glm.simpleCheckGLMGrid(self, glm, None, allowFailWarning=True, **kwargs)
# gets the failed to converge, here, after we see it in the browser too
x = re.compile("[Ff]ailed")
if warnings:
for w in warnings:
if (re.search(x,w)):
# first
if emsg is None: emsg = w
print w
if emsg: break
if not h2o.browse_disable:
h2b.browseJsonHistoryAsUrlLastMatch("GLMGridProgress")
time.sleep(5)
# gets the failed to converge, here, after we see it in the browser too
if emsg is not None:
raise Exception(emsg)
def test_parse_nflx_loop_s3n_hdfs(self):
DO_GLM = True
DO_GLMGRID = False
USE_S3 = False
noPoll = False
benchmarkLogging = ['jstack', 'iostats']
benchmarkLogging = ['iostats']
benchmarkLogging = []
# typical size of the michal files
avgMichalSize = 116561140
avgSynSize = 4020000
synSize = 183
csvFilenameList = [
(["manyfiles-nflx-gz"], "*file_1[0-9][0-9].dat.gz",
"file_100_A.dat.gz", 100 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[1-2][0-5][0-9].dat.gz",
"file_120_A.dat.gz", 120 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[1-2][0-6][0-9].dat.gz",
"file_140_A.dat.gz", 140 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[1-2][0-7][0-9].dat.gz",
"file_160_A.dat.gz", 160 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[1-2][0-8][0-9].dat.gz",
"file_180_A.dat.gz", 180 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[12][0-9][0-9].dat.gz",
"file_200_A.dat.gz", 200 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[123][0-9][0-9].dat.gz",
"file_300_A.dat.gz", 300 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[123][0-9][0-9].dat.gz",
"file_300_B.dat.gz", 300 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_[123][0-9][0-9].dat.gz",
"file_300_C.dat.gz", 300 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_1.dat.gz", "file_1.dat.gz",
1 * avgMichalSize, 300),
(["manyfiles-nflx-gz"], "*file_[2][0-9].dat.gz", "file_10.dat.gz",
10 * avgMichalSize, 700),
(["manyfiles-nflx-gz"], "*file_[34][0-9].dat.gz", "file_20.dat.gz",
20 * avgMichalSize, 900),
(["manyfiles-nflx-gz"], "*file_[5-9][0-9].dat.gz",
"file_50_A.dat.gz", 50 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_1[0-4][0-9].dat.gz",
"file_50_B.dat.gz", 50 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_1[0-9][0-9].dat.gz",
"file_100_A.dat.gz", 100 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "*file_2[0-9][0-9].dat.gz",
"file_100_B.dat.gz", 100 * avgMichalSize, 3600),
# beware: the files should be non-overlapping sequentially if noPoll is used, to avoid deleting keys in use
(["A-800-manyfiles-nflx-gz"], "*file_[0-9]*.dat.gz",
"file_A_200_x55.dat.gz", 200 * (avgMichalSize / 2), 7200),
(["A-800-manyfiles-nflx-gz",
"B-800-manyfiles-nflx-gz"], "*file_[0-9]*.dat.gz",
"file_A_400_x55.dat.gz", 400 * (avgMichalSize / 2), 7200),
([
"A-800-manyfiles-nflx-gz", "B-800-manyfiles-nflx-gz",
"C-800-manyfiles-nflx-gz", "D-800-manyfiles-nflx-gz"
], "*file_[0-9]*.dat.gz", "file_A_800_x55.dat.gz",
800 * (avgMichalSize / 2), 7200),
]
print "Using the -.gz files from s3"
# want just s3n://home-0xdiag-datasets/manyfiles-nflx-gz/file_1.dat.gz
# split out the pattern match and the filename used for the hex
trialMax = 1
pollTimeoutSecs = 180
retryDelaySecs = 10
# use i to forward reference in the list, so we can do multiple outstanding parses below
for i, (csvFolderList, csvFilepattern, csvFilename, totalBytes,
timeoutSecs) in enumerate(csvFilenameList):
bucket = "home-0xdiag-datasets"
## for tryHeap in [54, 28]:
h2oPerNode = 1
# h1.4xlarge 60.5GB dram
for tryHeap in [28]:
if USE_S3:
protocol = "s3"
else:
protocol = "s3n"
print "\n", tryHeap, "GB heap,", h2oPerNode, "jvm per host, import", protocol, "then parse"
# jea = "-XX:+UseParNewGC -XX:+UseConcMarkSweepGC"
# jea = "-Dh2o.find-ByteBuffer-leaks=true"
h2o_hosts.build_cloud_with_hosts(
h2oPerNode,
java_heap_GB=tryHeap,
# java_extra_args=jea,
enable_benchmark_log=True,
timeoutSecs=120,
retryDelaySecs=10,
# all hdfs info is done thru the hdfs_config michal's ec2 config sets up?
# this is for our amazon ec hdfs
# see https://github.com/0xdata/h2o/wiki/H2O-and-s3n
hdfs_name_node='10.78.14.235:9000',
hdfs_version='0.20.2')
# don't raise exception if we find something bad in h2o stdout/stderr?
h2o.nodes[0].sandbox_ignore_errors = True
for trial in range(trialMax):
# import a list of folders, one at a time (hdfs import can't take pattern match
# want to be able to parse 800 files, but only 200 per folder. Don't want to import the full bucket
# too slow
for csvFolder in csvFolderList:
if USE_S3:
URI = protocol + "://" + bucket + "/" + csvFolder + "/"
else:
URI = protocol + "://" + bucket + "/" + csvFolder + "/"
# since we delete the key, we have to re-import every iteration, to get it again
# s3n URI thru HDFS is not typical.
if USE_S3:
importResult = h2o.nodes[0].import_s3(bucket)
else:
importResult = h2o.nodes[0].import_hdfs(URI)
foundKeys = 0
for s in importResult['succeeded']:
# just print the first tile
# if 'nflx' in key and 'file_1.dat.gz' in key:
if csvFilepattern in s['key']:
# should be s3n://home-0xdiag-datasets/manyfiles-nflx-gz/file_1.dat.gz
print "example file we'll use:", s['key']
break
else:
pass
foundKeys += 1
### print "s3nFullList:", h2o.dump_json(s3nFullList)
# error if none?
self.assertGreater(
foundKeys, 8,
"Didn't see more than 8 files in s3n?")
s3nKey = csvFilepattern
key2 = csvFilename + "_" + str(trial) + ".hex"
print "Loading", protocol, "key:", s3nKey, "to", key2
start = time.time()
parseKey = h2o.nodes[0].parse(
s3nKey,
key2,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
noPoll=noPoll,
benchmarkLogging=benchmarkLogging)
if noPoll:
if (i + 1) < len(csvFilenameList):
time.sleep(1)
h2o.check_sandbox_for_errors()
(csvFilepattern, csvFilename, totalBytes2,
timeoutSecs) = csvFilenameList[i + 1]
s3nKey = URI + csvFilepattern
key2 = csvFilename + "_" + str(trial) + ".hex"
print "Loading", protocol, "key:", s3nKey, "to", key2
parse2Key = h2o.nodes[0].parse(
s3nKey,
key2,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
noPoll=noPoll,
benchmarkLogging=benchmarkLogging)
if (i + 2) < len(csvFilenameList):
time.sleep(1)
h2o.check_sandbox_for_errors()
(csvFilepattern, csvFilename, totalBytes3,
timeoutSecs) = csvFilenameList[i + 2]
s3nKey = URI + csvFilepattern
key2 = csvFilename + "_" + str(trial) + ".hex"
print "Loading", protocol, "key:", s3nKey, "to", key2
parse3Key = h2o.nodes[0].parse(
s3nKey,
key2,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
noPoll=noPoll,
benchmarkLogging=benchmarkLogging)
elapsed = time.time() - start
print s3nKey, 'parse time:', parseKey['response']['time']
print "parse result:", parseKey['destination_key']
print "Parse #", trial, "completed in", "%6.2f" % elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# print stats on all three if noPoll
if noPoll:
# does it take a little while to show up in Jobs, from where we issued the parse?
time.sleep(2)
# FIX! use the last (biggest?) timeoutSecs? maybe should increase since parallel
h2o_jobs.pollWaitJobs(
pattern=csvFilename,
timeoutSecs=timeoutSecs,
benchmarkLogging=benchmarkLogging)
# for getting the MB/sec closer to 'right'
totalBytes += totalBytes2 + totalBytes3
elapsed = time.time() - start
h2o.check_sandbox_for_errors()
if totalBytes is not None:
fileMBS = (totalBytes / 1e6) / elapsed
l = '{:d} jvms, {:d}GB heap, {:s} {:s} {:6.2f} MB/sec for {:6.2f} secs'.format(
len(h2o.nodes), tryHeap, csvFilepattern,
csvFilename, fileMBS, elapsed)
print l
h2o.cloudPerfH2O.message(l)
y = 378
if not noPoll:
x = h2o_glm.goodXFromColumnInfo(
y,
key=parseKey['destination_key'],
timeoutSecs=300)
#**********************************************************************************
# Do GLM too
# Argument case error: Value 0.0 is not between 12.0 and 9987.0 (inclusive)
if DO_GLM or DO_GLMGRID:
# these are all the columns that are enums in the dataset...too many for GLM!
x = range(542) # don't include the output column
# remove the output too! (378)
for i in [
3, 4, 5, 6, 7, 8, 9, 10, 11, 14, 16, 17, 18,
19, 20, 424, 425, 426, 540, 541, y
]:
x.remove(i)
x = ",".join(map(str, x))
if DO_GLM:
algo = 'GLM'
GLMkwargs = {
'x': x,
'y': y,
'case': 15,
'case_mode': '>',
'family': 'binomial',
'max_iter': 10,
'n_folds': 1,
'alpha': 0.2,
'lambda': 1e-5
}
start = time.time()
glm = h2o_cmd.runGLMOnly(
parseKey=parseKey,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
benchmarkLogging=benchmarkLogging,
**GLMkwargs)
elapsed = time.time() - start
h2o_glm.simpleCheckGLM(self, glm, None,
**GLMkwargs)
else:
algo = 'GLMGrid'
GLMkwargs = {
'x': x,
'y': y,
'case': 15,
'case_mode': '>',
'family': 'binomial',
'max_iter': 10,
'n_folds': 1,
'beta_epsilon': 1e-4,
'lambda': '1e-4',
'alpha': '0,0.5',
'thresholds': '0.5'
}
start = time.time()
glm = h2o_cmd.runGLMGridOnly(
parseKey=parseKey,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
benchmarkLogging=benchmarkLogging,
**GLMkwargs)
elapsed = time.time() - start
h2o_glm.simpleCheckGLMGrid(self, glm, None,
**GLMkwargs)
h2o.check_sandbox_for_errors()
l = '{:d} jvms, {:d}GB heap, {:s} {:s} {:s} {:6.2f} secs'.format(
len(h2o.nodes), tryHeap, algo, csvFilepattern,
csvFilename, elapsed)
print l
h2o.cloudPerfH2O.message(l)
#**********************************************************************************
print "Deleting key in H2O so we get it from S3 (if ec2) or nfs again.", \
"Otherwise it would just parse the cached key."
### storeView = h2o.nodes[0].store_view()
### print "storeView:", h2o.dump_json(storeView)
# "key": "s3n://home-0xdiag-datasets/manyfiles-nflx-gz/file_84.dat.gz"
# have to do the pattern match ourself, to figure out what keys to delete
# we're deleting the keys in the initial import. We leave the keys we created
# by the parse. We use unique dest keys for those, so no worries.
# Leaving them is good because things fill up! (spill)
h2o_cmd.checkKeyDistribution()
h2o_cmd.deleteCsvKey(csvFilename, importResult)
h2o.tear_down_cloud()
# sticky ports? wait a bit.
print "Waiting 30 secs before building cloud again (sticky ports?)"
time.sleep(30)
def test_parse_nflx_loop_s3n_hdfs(self):
DO_GLM = True
DO_GLMGRID = False
USE_S3 = False
noPoll = False
benchmarkLogging = ['jstack','iostats']
benchmarkLogging = ['iostats']
benchmarkLogging = []
# typical size of the michal files
avgMichalSize = 116561140
avgSynSize = 4020000
synSize = 183
csvFilenameList = [
(["manyfiles-nflx-gz"], "file_1[0-9][0-9].dat.gz", "file_100_A.dat.gz", 100 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[1-2][0-5][0-9].dat.gz", "file_120_A.dat.gz", 120 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[1-2][0-6][0-9].dat.gz", "file_140_A.dat.gz", 140 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[1-2][0-7][0-9].dat.gz", "file_160_A.dat.gz", 160 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[1-2][0-8][0-9].dat.gz", "file_180_A.dat.gz", 180 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[12][0-9][0-9].dat.gz", "file_200_A.dat.gz", 200 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[123][0-9][0-9].dat.gz", "file_300_A.dat.gz", 300 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[123][0-9][0-9].dat.gz", "file_300_B.dat.gz", 300 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_[123][0-9][0-9].dat.gz", "file_300_C.dat.gz", 300 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_1.dat.gz", "file_1.dat.gz", 1 * avgMichalSize, 300),
(["manyfiles-nflx-gz"], "file_[2][0-9].dat.gz", "file_10.dat.gz", 10 * avgMichalSize, 700),
(["manyfiles-nflx-gz"], "file_[34][0-9].dat.gz", "file_20.dat.gz", 20 * avgMichalSize, 900),
(["manyfiles-nflx-gz"], "file_[5-9][0-9].dat.gz", "file_50_A.dat.gz", 50 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_1[0-4][0-9].dat.gz", "file_50_B.dat.gz", 50 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_1[0-9][0-9].dat.gz", "file_100_A.dat.gz", 100 * avgMichalSize, 3600),
(["manyfiles-nflx-gz"], "file_2[0-9][0-9].dat.gz", "file_100_B.dat.gz", 100 * avgMichalSize, 3600),
# beware: the files should be non-overlapping sequentially if noPoll is used, to avoid deleting keys in use
(["A-800-manyfiles-nflx-gz"],
"*file_[0-9]*.dat.gz", "file_A_200_x55.dat.gz", 200 * (avgMichalSize/2), 7200),
(["A-800-manyfiles-nflx-gz", "B-800-manyfiles-nflx-gz"],
"*file_[0-9]*.dat.gz", "file_A_400_x55.dat.gz", 400 * (avgMichalSize/2), 7200),
(["A-800-manyfiles-nflx-gz", "B-800-manyfiles-nflx-gz", "C-800-manyfiles-nflx-gz", "D-800-manyfiles-nflx-gz"],
"*file_[0-9]*.dat.gz", "file_A_800_x55.dat.gz", 800 * (avgMichalSize/2), 7200),
]
print "Using the -.gz files from s3"
# want just s3n://home-0xdiag-datasets/manyfiles-nflx-gz/file_1.dat.gz
# split out the pattern match and the filename used for the hex
trialMax = 1
pollTimeoutSecs = 180
retryDelaySecs = 10
# use i to forward reference in the list, so we can do multiple outstanding parses below
for i, (csvFolderList, csvFilepattern, csvFilename, totalBytes, timeoutSecs) in enumerate(csvFilenameList):
bucket = "home-0xdiag-datasets"
## for tryHeap in [54, 28]:
h2oPerNode = 1
# h1.4xlarge 60.5GB dram
for tryHeap in [28]:
if USE_S3:
protocol = "s3"
else:
protocol = "s3n"
print "\n", tryHeap,"GB heap,", h2oPerNode, "jvm per host, import", protocol, "then parse"
# jea = "-XX:+UseParNewGC -XX:+UseConcMarkSweepGC"
# jea = "-Dh2o.find-ByteBuffer-leaks=true"
h2o_hosts.build_cloud_with_hosts(h2oPerNode, java_heap_GB=tryHeap,
# java_extra_args=jea,
enable_benchmark_log=True, timeoutSecs=120, retryDelaySecs=10,
# all hdfs info is done thru the hdfs_config michal's ec2 config sets up?
# this is for our amazon ec hdfs
# see https://github.com/0xdata/h2o/wiki/H2O-and-s3n
hdfs_name_node='10.78.14.235:9000',
hdfs_version='0.20.2')
# don't raise exception if we find something bad in h2o stdout/stderr?
h2o.nodes[0].sandbox_ignore_errors = True
for trial in range(trialMax):
# import a list of folders, one at a time (hdfs import can't take pattern match
# want to be able to parse 800 files, but only 200 per folder. Don't want to import the full bucket
# too slow
for csvFolder in csvFolderList:
if USE_S3:
URI = protocol + "://" + bucket + "/" + csvFolder + "/"
else:
URI = protocol + "://" + bucket + "/" + csvFolder + "/"
# since we delete the key, we have to re-import every iteration, to get it again
# s3n URI thru HDFS is not typical.
if USE_S3:
importResult = h2o.nodes[0].import_s3(bucket)
else:
importResult = h2o.nodes[0].import_hdfs(URI)
foundKeys = 0
for s in importResult['succeeded']:
# just print the first tile
# if 'nflx' in key and 'file_1.dat.gz' in key:
if csvFilepattern in s['key']:
# should be s3n://home-0xdiag-datasets/manyfiles-nflx-gz/file_1.dat.gz
print "example file we'll use:", s['key']
break
else:
pass
foundKeys += 1
### print "s3nFullList:", h2o.dump_json(s3nFullList)
# error if none?
self.assertGreater(foundKeys,8,"Didn't see more than 8 files in s3n?")
s3nKey = csvFilepattern
key2 = csvFilename + "_" + str(trial) + ".hex"
print "Loading", protocol, "key:", s3nKey, "to", key2
start = time.time()
parseKey = h2o.nodes[0].parse(s3nKey, key2,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
noPoll=noPoll,
benchmarkLogging=benchmarkLogging)
if noPoll:
if (i+1) < len(csvFilenameList):
time.sleep(1)
h2o.check_sandbox_for_errors()
(csvFilepattern, csvFilename, totalBytes2, timeoutSecs) = csvFilenameList[i+1]
s3nKey = URI + csvFilepattern
key2 = csvFilename + "_" + str(trial) + ".hex"
print "Loading", protocol, "key:", s3nKey, "to", key2
parse2Key = h2o.nodes[0].parse(s3nKey, key2,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
noPoll=noPoll,
benchmarkLogging=benchmarkLogging)
if (i+2) < len(csvFilenameList):
time.sleep(1)
h2o.check_sandbox_for_errors()
(csvFilepattern, csvFilename, totalBytes3, timeoutSecs) = csvFilenameList[i+2]
s3nKey = URI + csvFilepattern
key2 = csvFilename + "_" + str(trial) + ".hex"
print "Loading", protocol, "key:", s3nKey, "to", key2
parse3Key = h2o.nodes[0].parse(s3nKey, key2,
timeoutSecs=timeoutSecs,
retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
noPoll=noPoll,
benchmarkLogging=benchmarkLogging)
elapsed = time.time() - start
print s3nKey, 'parse time:', parseKey['response']['time']
print "parse result:", parseKey['destination_key']
print "Parse #", trial, "completed in", "%6.2f" % elapsed, "seconds.", \
"%d pct. of timeout" % ((elapsed*100)/timeoutSecs)
# print stats on all three if noPoll
if noPoll:
# does it take a little while to show up in Jobs, from where we issued the parse?
time.sleep(2)
# FIX! use the last (biggest?) timeoutSecs? maybe should increase since parallel
h2o_jobs.pollWaitJobs(pattern=csvFilename,
timeoutSecs=timeoutSecs, benchmarkLogging=benchmarkLogging)
# for getting the MB/sec closer to 'right'
totalBytes += totalBytes2 + totalBytes3
elapsed = time.time() - start
h2o.check_sandbox_for_errors()
if totalBytes is not None:
fileMBS = (totalBytes/1e6)/elapsed
l = '{:d} jvms, {:d}GB heap, {:s} {:s} {:6.2f} MB/sec for {:6.2f} secs'.format(
len(h2o.nodes), tryHeap, csvFilepattern, csvFilename, fileMBS, elapsed)
print l
h2o.cloudPerfH2O.message(l)
y = 378
if not noPoll:
x = h2o_glm.goodXFromColumnInfo(y, key=parseKey['destination_key'], timeoutSecs=300)
#**********************************************************************************
# Do GLM too
# Argument case error: Value 0.0 is not between 12.0 and 9987.0 (inclusive)
if DO_GLM or DO_GLMGRID:
# these are all the columns that are enums in the dataset...too many for GLM!
x = range(542) # don't include the output column
# remove the output too! (378)
for i in [3, 4, 5, 6, 7, 8, 9, 10, 11, 14, 16, 17, 18, 19, 20, 424, 425, 426, 540, 541, y]:
x.remove(i)
x = ",".join(map(str,x))
if DO_GLM:
algo = 'GLM'
GLMkwargs = {'x': x, 'y': y, 'case': 15, 'case_mode': '>', 'family': 'binomial',
'max_iter': 10, 'n_folds': 1, 'alpha': 0.2, 'lambda': 1e-5}
start = time.time()
glm = h2o_cmd.runGLMOnly(parseKey=parseKey,
timeoutSecs=timeoutSecs, retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
benchmarkLogging=benchmarkLogging, **GLMkwargs)
elapsed = time.time() - start
h2o_glm.simpleCheckGLM(self, glm, None, **GLMkwargs)
else:
algo = 'GLMGrid'
GLMkwargs = {'x': x, 'y': y, 'case': 15, 'case_mode': '>', 'family': 'binomial',
'max_iter': 10, 'n_folds': 1, 'beta_epsilon': 1e-4,
'lambda': '1e-4',
'alpha': '0,0.5',
'thresholds': '0.5'
}
start = time.time()
glm = h2o_cmd.runGLMGridOnly(parseKey=parseKey,
timeoutSecs=timeoutSecs, retryDelaySecs=retryDelaySecs,
pollTimeoutSecs=pollTimeoutSecs,
benchmarkLogging=benchmarkLogging, **GLMkwargs)
elapsed = time.time() - start
h2o_glm.simpleCheckGLMGrid(self, glm, None, **GLMkwargs)
h2o.check_sandbox_for_errors()
l = '{:d} jvms, {:d}GB heap, {:s} {:s} {:s} {:6.2f} secs'.format(
len(h2o.nodes), tryHeap, algo, csvFilepattern, csvFilename, elapsed)
print l
h2o.cloudPerfH2O.message(l)
#**********************************************************************************
print "Deleting key in H2O so we get it from S3 (if ec2) or nfs again.", \
"Otherwise it would just parse the cached key."
### storeView = h2o.nodes[0].store_view()
### print "storeView:", h2o.dump_json(storeView)
# "key": "s3n://home-0xdiag-datasets/manyfiles-nflx-gz/file_84.dat.gz"
# have to do the pattern match ourself, to figure out what keys to delete
# we're deleting the keys in the initial import. We leave the keys we created
# by the parse. We use unique dest keys for those, so no worries.
# Leaving them is good because things fill up! (spill)
h2o_cmd.checkKeyDistribution()
h2o_cmd.deleteCsvKey(csvFilename, importResult)
h2o.tear_down_cloud()
# sticky ports? wait a bit.
print "Waiting 30 secs before building cloud again (sticky ports?)"
time.sleep(30)
def test_GLM_convergence_1(self):
SYNDATASETS_DIR = h2o.make_syn_dir()
tryList = [
(100, 50, 'cD', 300),
(100, 100, 'cE', 300),
(100, 200, 'cF', 300),
(100, 300, 'cG', 300),
(100, 400, 'cH', 300),
(100, 500, 'cI', 300),
]
### h2b.browseTheCloud()
lenNodes = len(h2o.nodes)
USEKNOWNFAILURE = True
for (rowCount, colCount, key2, timeoutSecs) in tryList:
SEEDPERFILE = random.randint(0, sys.maxint)
csvFilename = 'syn_%s_%sx%s.csv' % (SEEDPERFILE,rowCount,colCount)
csvPathname = SYNDATASETS_DIR + '/' + csvFilename
print "\nCreating random", csvPathname
write_syn_dataset(csvPathname, rowCount, colCount, SEEDPERFILE)
if USEKNOWNFAILURE:
csvFilename = 'failtoconverge_100x50.csv'
csvPathname = h2o.find_file('smalldata/logreg/' + csvFilename)
parseKey = h2o_cmd.parseFile(None, csvPathname, key2=key2, timeoutSecs=10)
print csvFilename, 'parse time:', parseKey['response']['time']
print "Parse result['destination_key']:", parseKey['destination_key']
inspect = h2o_cmd.runInspect(None, parseKey['destination_key'])
print "\n" + csvFilename
y = colCount
kwargs = {
'max_iter': 10,
'weight': 1.0,
'link': 'familyDefault',
'n_folds': 2,
'beta_epsilon': 1e-4,
'lambda': '1e-8:1e-3:1e2',
'alpha': '0,0.5,.75',
'thresholds': '0,1,0.2'
}
if USEKNOWNFAILURE:
kwargs['y'] = 50
else:
kwargs['y'] = y
emsg = None
for i in range(2):
start = time.time()
# get rid of the Jstack polling
glm = h2o_cmd.runGLMGridOnly(parseKey=parseKey, timeoutSecs=timeoutSecs, **kwargs)
print 'glm #', i, 'end on', csvPathname, 'took', time.time() - start, 'seconds'
# we can pass the warning, without stopping in the test, so we can
# redo it in the browser for comparison
warnings = h2o_glm.simpleCheckGLMGrid(self, glm, None, allowFailWarning=True, **kwargs)
# gets the failed to converge, here, after we see it in the browser too
x = re.compile("[Ff]ailed")
if warnings:
for w in warnings:
if (re.search(x,w)):
# first
if emsg is None: emsg = w
print w
if emsg: break
if not h2o.browse_disable:
h2b.browseJsonHistoryAsUrlLastMatch("GLMGridProgress")
time.sleep(5)
# gets the failed to converge, here, after we see it in the browser too
if emsg is not None:
raise Exception(emsg)